Next: The Algorithm Up: Learning Agents for Previous: The proposed approach

Comparison with previous approaches

A number of differences can be identified on comparing the proposed approach with previous approaches to the personalized information filtering problem. The differences are discussed below.

It is difficult to model exploration using rule-based systems which try to detect patterns in the user's behavior. For example, INFOSCOPE learns using rule based systems which remember interesting topics covered in the past. New recommendations of topics are made to the user based on recency, frequency and spacing of past topics [9]. The disadvantage of such an approach is that it is constrained to making recommendations of topics which lie within the realm of user's past interests. On the other hand, we explicitly model exploration. The information filtering agent searches new domains for information which could be of potential interest to the user. It is quite possible that user may not have seen the topic before.

Our approach is quite similar to the work described in Yang and Korfhage [47]. They evolve a population of query individuals, we evolve a population of profile individuals. However, they assume that user interests are fixed and strive towards convergence. We assume dynamic user interests and our goal is to continually adapt. Another difference is that each profile individual in our system learns within it's lifetime taking advantage of the Baldwin effect described above. This is not the case in the other system.

Our approach differs in many respects from Information Intake Filtering (IIF) [2] mentioned earlier. The first difference is that IIF is a subset of the personalized information filtering problem as defined in this thesis. By assuming an in-basket, it scales down the problem considerably to that of prioritizing articles that have already undergone one level of filtering before reaching the in-basket. The feature space in IIF depends on the size of the intake and therefore the in-basket cannot be scaled up to the universe of documents. Another difference is that IIF implicitly assumes that user interests stay fixed once learned. Furthermore, it does not explore newer information domains, which is the function of the mutation operator in Newt. The use of an economic system with GA for assigning payoffs is interesting, however, and deserves further research.

Doppelgänger [35] is a user modeling system briefly described above. The most important distinction with our approach to user modeling and is that Doppelgänger is application independent. Calling it a user modeling shell would be more appropriate to distinguish it from user models that are domain dependent. Since Doppelgänger solves a more general problem, there cannot be a direct comparison with the domain dependent user model acquired in Newt. The following comparison is only in the context of the information filtering problem being addressed in this thesis. The advantage of an application independent user model is that it is able to identify patterns in user behavior that are not performed while reading news but is important nonetheless to the IF problem. An example of such an inference is that the user reads business news immediately after login and weather news just before logout. The disadvantage is that a lot of this information might be irrelevant, perhaps even counter-productive to filtering. By comparison, the representation of profiles in Newt has been strongly motivated by the domain. It has been found to be a sufficient user model for keyword based filtering. Another difference is that some of the inferencing techniques used in Doppelgänger make it difficult to find reasons for the inference. The relatively simpler keyword based model can provide reasonable explanations for selecting or rejecting documents. There are further differences in the learning techniques used. Due to the nature of the data collected (locations, states, login data, etc.), Doppelgänger uses Markov models, time series techniques and clustering techniques. In our approach, we are interested in an efficient parallel search and have chosen genetic algorithms.



Next: The Algorithm Up: Learning Agents for Previous: The proposed approach


MIT Media Lab - Autonomous Agents Group - agentmaster@media.mit.edu